Extracting Information from Old and Scanned Engineering Drawings of Existing Buildings for the Creation of Digital Building Models

As a 3D model is considered the core element of building information modeling (BIM) applications, current research focuses on optimizing and automating the creation of BIM models, especially for existing buildings. In this regard, engineering drawings are the most common data source for extracting geometrical information. With the recent emersion of artificial intelligence (AI) applications and the availability of advanced algorithms, researchers have utilized computer vision approaches to process engineering drawings. However, the data used in previous studies is rarely derived from actual projects and is, in many cases, noise-free. Such datasets overlook the need to create BIM models for existing structures, which account for most buildings. To address this issue, the current paper describes an approach for extracting information from engineering drawings that originate from an existing infrastructure project in Germany. More precisely, the drawings depict the buildings located on both sides of a subway line. The scanned drawings are old and, in some cases, damaged. Moreover, such data collections are usually submitted or archived without a clear structure or logical naming convention, making data organization time-consuming and labor-intensive. The presented approach divides the extraction of information into two levels. The first level classifies the drawings according to the main content, such as different views of the building. The second level collects information from the first level’s identified views concerning structural and architectural aspects such as staircases, rooms, and openings. As the approach follows the idea of a data-centric AI, the pipeline includes an intensive exploration of the data, as well as its preprocessing, augmentation, and handling of damages in the drawings. The described approach is tested across multiple datasets and shows promising results. This study may represent an important step toward the automatic creation of digital twins for existing buildings.